1 /* Copyright (c) 2016 Facebook 2 * 3 * This program is free software; you can redistribute it and/or 4 * modify it under the terms of version 2 of the GNU General Public 5 * License as published by the Free Software Foundation. 6 */ 7 #define _GNU_SOURCE 8 #include <sched.h> 9 #include <stdio.h> 10 #include <sys/types.h> 11 #include <asm/unistd.h> 12 #include <unistd.h> 13 #include <assert.h> 14 #include <sys/wait.h> 15 #include <stdlib.h> 16 #include <signal.h> 17 #include <linux/bpf.h> 18 #include <string.h> 19 #include <time.h> 20 #include <sys/resource.h> 21 #include <arpa/inet.h> 22 #include <errno.h> 23 24 #include "libbpf.h" 25 #include "bpf_load.h" 26 27 #define TEST_BIT(t) (1U << (t)) 28 #define MAX_NR_CPUS 1024 29 30 static __u64 time_get_ns(void) 31 { 32 struct timespec ts; 33 34 clock_gettime(CLOCK_MONOTONIC, &ts); 35 return ts.tv_sec * 1000000000ull + ts.tv_nsec; 36 } 37 38 enum test_type { 39 HASH_PREALLOC, 40 PERCPU_HASH_PREALLOC, 41 HASH_KMALLOC, 42 PERCPU_HASH_KMALLOC, 43 LRU_HASH_PREALLOC, 44 NOCOMMON_LRU_HASH_PREALLOC, 45 LPM_KMALLOC, 46 HASH_LOOKUP, 47 ARRAY_LOOKUP, 48 INNER_LRU_HASH_PREALLOC, 49 LRU_HASH_LOOKUP, 50 NR_TESTS, 51 }; 52 53 const char *test_map_names[NR_TESTS] = { 54 [HASH_PREALLOC] = "hash_map", 55 [PERCPU_HASH_PREALLOC] = "percpu_hash_map", 56 [HASH_KMALLOC] = "hash_map_alloc", 57 [PERCPU_HASH_KMALLOC] = "percpu_hash_map_alloc", 58 [LRU_HASH_PREALLOC] = "lru_hash_map", 59 [NOCOMMON_LRU_HASH_PREALLOC] = "nocommon_lru_hash_map", 60 [LPM_KMALLOC] = "lpm_trie_map_alloc", 61 [HASH_LOOKUP] = "hash_map", 62 [ARRAY_LOOKUP] = "array_map", 63 [INNER_LRU_HASH_PREALLOC] = "inner_lru_hash_map", 64 [LRU_HASH_LOOKUP] = "lru_hash_lookup_map", 65 }; 66 67 static int test_flags = ~0; 68 static uint32_t num_map_entries; 69 static uint32_t inner_lru_hash_size; 70 static int inner_lru_hash_idx = -1; 71 static int array_of_lru_hashs_idx = -1; 72 static int lru_hash_lookup_idx = -1; 73 static int lru_hash_lookup_test_entries = 32; 74 static uint32_t max_cnt = 1000000; 75 76 static int check_test_flags(enum test_type t) 77 { 78 return test_flags & TEST_BIT(t); 79 } 80 81 static void test_hash_prealloc(int cpu) 82 { 83 __u64 start_time; 84 int i; 85 86 start_time = time_get_ns(); 87 for (i = 0; i < max_cnt; i++) 88 syscall(__NR_getuid); 89 printf("%d:hash_map_perf pre-alloc %lld events per sec\n", 90 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time)); 91 } 92 93 static int pre_test_lru_hash_lookup(int tasks) 94 { 95 int fd = map_fd[lru_hash_lookup_idx]; 96 uint32_t key; 97 long val = 1; 98 int ret; 99 100 if (num_map_entries > lru_hash_lookup_test_entries) 101 lru_hash_lookup_test_entries = num_map_entries; 102 103 /* Populate the lru_hash_map for LRU_HASH_LOOKUP perf test. 104 * 105 * It is fine that the user requests for a map with 106 * num_map_entries < 32 and some of the later lru hash lookup 107 * may return not found. For LRU map, we are not interested 108 * in such small map performance. 109 */ 110 for (key = 0; key < lru_hash_lookup_test_entries; key++) { 111 ret = bpf_map_update_elem(fd, &key, &val, BPF_NOEXIST); 112 if (ret) 113 return ret; 114 } 115 116 return 0; 117 } 118 119 static void do_test_lru(enum test_type test, int cpu) 120 { 121 static int inner_lru_map_fds[MAX_NR_CPUS]; 122 123 struct sockaddr_in6 in6 = { .sin6_family = AF_INET6 }; 124 const char *test_name; 125 __u64 start_time; 126 int i, ret; 127 128 if (test == INNER_LRU_HASH_PREALLOC) { 129 int outer_fd = map_fd[array_of_lru_hashs_idx]; 130 unsigned int mycpu, mynode; 131 132 assert(cpu < MAX_NR_CPUS); 133 134 if (cpu) { 135 ret = syscall(__NR_getcpu, &mycpu, &mynode, NULL); 136 assert(!ret); 137 138 inner_lru_map_fds[cpu] = 139 bpf_create_map_node(BPF_MAP_TYPE_LRU_HASH, 140 sizeof(uint32_t), 141 sizeof(long), 142 inner_lru_hash_size, 0, 143 mynode); 144 if (inner_lru_map_fds[cpu] == -1) { 145 printf("cannot create BPF_MAP_TYPE_LRU_HASH %s(%d)\n", 146 strerror(errno), errno); 147 exit(1); 148 } 149 } else { 150 inner_lru_map_fds[cpu] = map_fd[inner_lru_hash_idx]; 151 } 152 153 ret = bpf_map_update_elem(outer_fd, &cpu, 154 &inner_lru_map_fds[cpu], 155 BPF_ANY); 156 if (ret) { 157 printf("cannot update ARRAY_OF_LRU_HASHS with key:%u. %s(%d)\n", 158 cpu, strerror(errno), errno); 159 exit(1); 160 } 161 } 162 163 in6.sin6_addr.s6_addr16[0] = 0xdead; 164 in6.sin6_addr.s6_addr16[1] = 0xbeef; 165 166 if (test == LRU_HASH_PREALLOC) { 167 test_name = "lru_hash_map_perf"; 168 in6.sin6_addr.s6_addr16[2] = 0; 169 } else if (test == NOCOMMON_LRU_HASH_PREALLOC) { 170 test_name = "nocommon_lru_hash_map_perf"; 171 in6.sin6_addr.s6_addr16[2] = 1; 172 } else if (test == INNER_LRU_HASH_PREALLOC) { 173 test_name = "inner_lru_hash_map_perf"; 174 in6.sin6_addr.s6_addr16[2] = 2; 175 } else if (test == LRU_HASH_LOOKUP) { 176 test_name = "lru_hash_lookup_perf"; 177 in6.sin6_addr.s6_addr16[2] = 3; 178 in6.sin6_addr.s6_addr32[3] = 0; 179 } else { 180 assert(0); 181 } 182 183 start_time = time_get_ns(); 184 for (i = 0; i < max_cnt; i++) { 185 ret = connect(-1, (const struct sockaddr *)&in6, sizeof(in6)); 186 assert(ret == -1 && errno == EBADF); 187 if (in6.sin6_addr.s6_addr32[3] < 188 lru_hash_lookup_test_entries - 32) 189 in6.sin6_addr.s6_addr32[3] += 32; 190 else 191 in6.sin6_addr.s6_addr32[3] = 0; 192 } 193 printf("%d:%s pre-alloc %lld events per sec\n", 194 cpu, test_name, 195 max_cnt * 1000000000ll / (time_get_ns() - start_time)); 196 } 197 198 static void test_lru_hash_prealloc(int cpu) 199 { 200 do_test_lru(LRU_HASH_PREALLOC, cpu); 201 } 202 203 static void test_nocommon_lru_hash_prealloc(int cpu) 204 { 205 do_test_lru(NOCOMMON_LRU_HASH_PREALLOC, cpu); 206 } 207 208 static void test_inner_lru_hash_prealloc(int cpu) 209 { 210 do_test_lru(INNER_LRU_HASH_PREALLOC, cpu); 211 } 212 213 static void test_lru_hash_lookup(int cpu) 214 { 215 do_test_lru(LRU_HASH_LOOKUP, cpu); 216 } 217 218 static void test_percpu_hash_prealloc(int cpu) 219 { 220 __u64 start_time; 221 int i; 222 223 start_time = time_get_ns(); 224 for (i = 0; i < max_cnt; i++) 225 syscall(__NR_geteuid); 226 printf("%d:percpu_hash_map_perf pre-alloc %lld events per sec\n", 227 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time)); 228 } 229 230 static void test_hash_kmalloc(int cpu) 231 { 232 __u64 start_time; 233 int i; 234 235 start_time = time_get_ns(); 236 for (i = 0; i < max_cnt; i++) 237 syscall(__NR_getgid); 238 printf("%d:hash_map_perf kmalloc %lld events per sec\n", 239 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time)); 240 } 241 242 static void test_percpu_hash_kmalloc(int cpu) 243 { 244 __u64 start_time; 245 int i; 246 247 start_time = time_get_ns(); 248 for (i = 0; i < max_cnt; i++) 249 syscall(__NR_getegid); 250 printf("%d:percpu_hash_map_perf kmalloc %lld events per sec\n", 251 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time)); 252 } 253 254 static void test_lpm_kmalloc(int cpu) 255 { 256 __u64 start_time; 257 int i; 258 259 start_time = time_get_ns(); 260 for (i = 0; i < max_cnt; i++) 261 syscall(__NR_gettid); 262 printf("%d:lpm_perf kmalloc %lld events per sec\n", 263 cpu, max_cnt * 1000000000ll / (time_get_ns() - start_time)); 264 } 265 266 static void test_hash_lookup(int cpu) 267 { 268 __u64 start_time; 269 int i; 270 271 start_time = time_get_ns(); 272 for (i = 0; i < max_cnt; i++) 273 syscall(__NR_getpgid, 0); 274 printf("%d:hash_lookup %lld lookups per sec\n", 275 cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time)); 276 } 277 278 static void test_array_lookup(int cpu) 279 { 280 __u64 start_time; 281 int i; 282 283 start_time = time_get_ns(); 284 for (i = 0; i < max_cnt; i++) 285 syscall(__NR_getpgrp, 0); 286 printf("%d:array_lookup %lld lookups per sec\n", 287 cpu, max_cnt * 1000000000ll * 64 / (time_get_ns() - start_time)); 288 } 289 290 typedef int (*pre_test_func)(int tasks); 291 const pre_test_func pre_test_funcs[] = { 292 [LRU_HASH_LOOKUP] = pre_test_lru_hash_lookup, 293 }; 294 295 typedef void (*test_func)(int cpu); 296 const test_func test_funcs[] = { 297 [HASH_PREALLOC] = test_hash_prealloc, 298 [PERCPU_HASH_PREALLOC] = test_percpu_hash_prealloc, 299 [HASH_KMALLOC] = test_hash_kmalloc, 300 [PERCPU_HASH_KMALLOC] = test_percpu_hash_kmalloc, 301 [LRU_HASH_PREALLOC] = test_lru_hash_prealloc, 302 [NOCOMMON_LRU_HASH_PREALLOC] = test_nocommon_lru_hash_prealloc, 303 [LPM_KMALLOC] = test_lpm_kmalloc, 304 [HASH_LOOKUP] = test_hash_lookup, 305 [ARRAY_LOOKUP] = test_array_lookup, 306 [INNER_LRU_HASH_PREALLOC] = test_inner_lru_hash_prealloc, 307 [LRU_HASH_LOOKUP] = test_lru_hash_lookup, 308 }; 309 310 static int pre_test(int tasks) 311 { 312 int i; 313 314 for (i = 0; i < NR_TESTS; i++) { 315 if (pre_test_funcs[i] && check_test_flags(i)) { 316 int ret = pre_test_funcs[i](tasks); 317 318 if (ret) 319 return ret; 320 } 321 } 322 323 return 0; 324 } 325 326 static void loop(int cpu) 327 { 328 cpu_set_t cpuset; 329 int i; 330 331 CPU_ZERO(&cpuset); 332 CPU_SET(cpu, &cpuset); 333 sched_setaffinity(0, sizeof(cpuset), &cpuset); 334 335 for (i = 0; i < NR_TESTS; i++) { 336 if (check_test_flags(i)) 337 test_funcs[i](cpu); 338 } 339 } 340 341 static void run_perf_test(int tasks) 342 { 343 pid_t pid[tasks]; 344 int i; 345 346 assert(!pre_test(tasks)); 347 348 for (i = 0; i < tasks; i++) { 349 pid[i] = fork(); 350 if (pid[i] == 0) { 351 loop(i); 352 exit(0); 353 } else if (pid[i] == -1) { 354 printf("couldn't spawn #%d process\n", i); 355 exit(1); 356 } 357 } 358 for (i = 0; i < tasks; i++) { 359 int status; 360 361 assert(waitpid(pid[i], &status, 0) == pid[i]); 362 assert(status == 0); 363 } 364 } 365 366 static void fill_lpm_trie(void) 367 { 368 struct bpf_lpm_trie_key *key; 369 unsigned long value = 0; 370 unsigned int i; 371 int r; 372 373 key = alloca(sizeof(*key) + 4); 374 key->prefixlen = 32; 375 376 for (i = 0; i < 512; ++i) { 377 key->prefixlen = rand() % 33; 378 key->data[0] = rand() & 0xff; 379 key->data[1] = rand() & 0xff; 380 key->data[2] = rand() & 0xff; 381 key->data[3] = rand() & 0xff; 382 r = bpf_map_update_elem(map_fd[6], key, &value, 0); 383 assert(!r); 384 } 385 386 key->prefixlen = 32; 387 key->data[0] = 192; 388 key->data[1] = 168; 389 key->data[2] = 0; 390 key->data[3] = 1; 391 value = 128; 392 393 r = bpf_map_update_elem(map_fd[6], key, &value, 0); 394 assert(!r); 395 } 396 397 static void fixup_map(struct bpf_map_data *map, int idx) 398 { 399 int i; 400 401 if (!strcmp("inner_lru_hash_map", map->name)) { 402 inner_lru_hash_idx = idx; 403 inner_lru_hash_size = map->def.max_entries; 404 } 405 406 if (!strcmp("array_of_lru_hashs", map->name)) { 407 if (inner_lru_hash_idx == -1) { 408 printf("inner_lru_hash_map must be defined before array_of_lru_hashs\n"); 409 exit(1); 410 } 411 map->def.inner_map_idx = inner_lru_hash_idx; 412 array_of_lru_hashs_idx = idx; 413 } 414 415 if (!strcmp("lru_hash_lookup_map", map->name)) 416 lru_hash_lookup_idx = idx; 417 418 if (num_map_entries <= 0) 419 return; 420 421 inner_lru_hash_size = num_map_entries; 422 423 /* Only change the max_entries for the enabled test(s) */ 424 for (i = 0; i < NR_TESTS; i++) { 425 if (!strcmp(test_map_names[i], map->name) && 426 (check_test_flags(i))) { 427 map->def.max_entries = num_map_entries; 428 } 429 } 430 } 431 432 int main(int argc, char **argv) 433 { 434 struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY}; 435 char filename[256]; 436 int num_cpu = 8; 437 438 snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]); 439 setrlimit(RLIMIT_MEMLOCK, &r); 440 441 if (argc > 1) 442 test_flags = atoi(argv[1]) ? : test_flags; 443 444 if (argc > 2) 445 num_cpu = atoi(argv[2]) ? : num_cpu; 446 447 if (argc > 3) 448 num_map_entries = atoi(argv[3]); 449 450 if (argc > 4) 451 max_cnt = atoi(argv[4]); 452 453 if (load_bpf_file_fixup_map(filename, fixup_map)) { 454 printf("%s", bpf_log_buf); 455 return 1; 456 } 457 458 fill_lpm_trie(); 459 460 run_perf_test(num_cpu); 461 462 return 0; 463 } 464